Download Models for a binary dependent variable

Economics 215, Fall 2011
Allin Cottrell
Models for a binary dependent variable
A binary dependent variable is one that can only take on values 0 or 1 at each observation; typically
it’s a coding of something qualitative (e.g. married versus not married, approved for a loan versus not
approved).
1
The Linear Probability Model (LPM)
In the LPM we estimate the standard linear model
y = Xβ + u
(1)
using OLS.
Under the unbiasedness assumption E(u|X) = 0 we see that E(y) = Xβ. But, since y is discrete, we
can also compute E(y) from first principles as the probability-weighted sum of its possible values:
E(y) = 1 × P (y = 1) + 0 × P (y = 0) = P (y = 1)
(2)
It follows that Xβ = P (y = 1). For convenience, we’ll refer to this common value as p (and the value at
each observation as pi ). Given an estimate of the parameter vector, β̂, the fitted value, Xi β̂, represents
the estimated probability that y takes on the value 1 at observation i.
This raises a first objection to the LPM: while the fitted values should be interpretable as probabilities,
there’s nothing in the OLS mechanism to ensure that 0 ≤ Xi β̂ ≤ 1 at all observations, i.e. nothing to
ensure that the fitted values are “legitimate” probabilities.
Now consider the error term in (1), namely,
u = y − Xβ = y − p
If yi = 1 then ui = 1 − pi , and if yi = 0 then ui = −pi . Still assuming E(u|X) = 0, the variance of ui
can be calculated as the probability-weighted sum of the squares of the possible values of u:
σi2 = E(u2i ) = (1 − pi )2 × pi + (−pi )2 × (1 − pi )
which simplifies to pi (1 − pi ). This means that if pi differs across observations depending on the
values of Xi (as it must if the model is to be any use), the variance of the error term is non-constant.
That is, the LPM is inherently heteroskedastic.
2
Maximum Likelihood estimation
The standard ML approach to the model with a binary dependent variable is to postulate a continuous
“latent variable”, y ∗ , such that
y ∗ = Xβ + u
(3)
where u follows some well-defined probability distribution. The observed dependent variable, y, is
then linked to the unobserved y ∗ via
yi =
(
1
if yi∗ > 0
0
if yi∗ ≤ 0
(4)
If the yi values represent the outcomes of decisions made by individuals (buy a car, emigrate, vote in
the Presidential election or whatever) then the latent variable can be interpreted as the perceived net
benefit or utility of the action—if the net benefit is positive, perform the action, otherwise don’t.
The distribution of the error term is generally assumed to be either standard normal (giving the probit
model) or the logistic distribution (logit model).
How can we express the probabilities of the two outcomes in this sort of model? Well, we have
P (yi = 1) = P (yi∗ > 0)
= P (Xi β + ui > 0)
= P (ui > −Xi β)
= P (ui < Xi β)
The last line works on condition that the distribution of u is symmetrical about zero (which is the case
for both probit and logit). For example, let Xi β = 2.15; then we’ll get yi∗ > 0, and hence yi = 1, so
long as ui > −2.15. But by symmetry, that has the same probability as ui < 2.15.
By the complementarity of probabilities, P (yi = 0) must equal 1 − P (yi = 1) = P (ui < 1 − Xi β).
The expression P (ui < Xi β) denotes the CDF of u evaluated at Xi β. In the probit model this is the
normal CDF, Φ(Xi β), and in the logit model it is the logistic CDF, Λ(Xi β) where Λ(x) = 1/(1 + e−x ).
We’ll use the symbol F (Xi β) to indicate either of these functions, depending on the context.
To obtain estimates for the model composed of equations (3) and (4), we employ an algorithm that
selects β̂ to as to maximize the joint likelihood of the sample data (as a practical matter we in fact
work with the log-likelihood). Each data point makes the following contribution to the log-likelihood:
`i = yi × log P (yi = 1) + (1 − yi ) × log P (yi = 0)
= yi log F (Xi β) + (1 − yi ) log F (1 − Xi β)
At each observation, either yi = 1, in which case the first term is non-zero and the second zero, or
yi = 0, in which case the reverse is true. The joint log-likelihood is:
`=
n
X
yi log F (Xi β) + (1 − yi ) log F (1 − Xi β)
i=1
This is what we’re asking the computer to maximize for us (with respect to β̂) when we call for probit
or logit estimates.
3
Reading probit and logit estimates
Some care is needed in interpreting the estimated coefficients (β̂) from these models. Note that these
do not represent the marginal effects of the independent variables on the probability that yi = 1.
Rather they are just the effects on the sum Xi β̂. To get the marginal effects we’re interested in we
need to find
∂F (Xi β̂)
∂Xi
for the nonlinear function F . These effects will not be constant. For convenience, gretl prints each
effect evaluated at the means of all the independent variables.
It often makes sense to compute a series of effects to get a fuller sense of what the model implies.
For example, consider a probit model using using a dataset from T. A. Mroz (gretl’s mroz87.gdt)
containing information on 753 women. The binary dependent variable, LFP, takes a value of 1 if the
woman participated in the labor force in 1975, otherwise 0. The dataset contains several covariates
2
which might plausibly influence a woman’s decision to seek paid employment: we’ll use KL6 (number
of children under the age of 6); WA, the woman’s age; WE, her education level; and MTR, the marginal tax
rate she faces.
Let’s see how we can find the probability that a woman is in the labor force for each case in KL6 = (0,
1, 2, 3). We need to choose values for the other elements of the X matrix, and the simplest thing to do
is set them at their sample means. The following gretl script will do the job.
open mroz87.gdt
# list of independent variables
list Xlist = KL6 WA WE MTR
# set variables other than KL6 to their sample means
matrix Xrow = { 1, 0, mean(WA), mean(WE), mean(MTR) }
matrix X = Xrow | Xrow | Xrow | Xrow
# set KL6 values from 0 to 3
matrix Kcol = { 0, 1, 2, 3 }’
X[,2] = Kcol
# check X
print X
# run probit
probit LFP 0 Xlist
# compute Phi(X\beta)
matrix P = Kcol ˜ cnorm(X*$coeff)
# give names to the columns of P (optional)
colnames(P, "KL6 P(LFP=1)")
print P
g1 <- gnuplot 2 1 --matrix=P --suppress --with-lines
g1.show
The output reads, in part:
X (4 x 5)
1.0000
1.0000
1.0000
1.0000
0.0000
1.0000
2.0000
3.0000
42.538
42.538
42.538
42.538
12.287
12.287
12.287
12.287
0.67886
0.67886
0.67886
0.67886
Model 1: Probit, using observations 1-753
Dependent variable: LFP
coefficient
std. error
t-ratio
slope
---------------------------------------------------------const
1.32275
0.705847
1.874
KL6
-0.852144
0.111001
-7.677
-0.334180
WA
-0.0347571
0.00673706
-5.159
-0.0136305
WE
0.105956
0.0240874
4.399
0.0415522
MTR
-1.11720
0.639522
-1.747
-0.438127
P (4 x 2)
KL6
0.0000
1.0000
2.0000
3.0000
P(LFP=1)
0.65088
0.32116
0.093988
0.015052
3
Thus we see that for the first child under 6, the probability of participating in the labor force drops
off from 0.65 to 0.32; for the second it falls to 0.094; and for the third, to 0.015. Not surprisingly, we
get a diminishing marginal impact, as shown in Figure 1.
0.7
0.6
P(LFP=1)
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
3
KL6
Figure 1: Probability of LFP as a function of KL6
Note that the Linear Probability Model, by contrast, imposes a constant marginal impact, which may
be unrealistic. If we apply OLS to the model specified above and repeat the calculation of the effect of
varying KL6 from 0 to 3 which holding the other regressors at their sample means, we get the following
rather unsatisfactory result:
KL6
0.0000
1.0000
2.0000
3.0000
P(LFP=1)
0.63928
0.34107
0.042847
-0.25537
4